Prompt Guide: Qwen-Image-2.1

Qwen-Image-2.1 is Alibaba's most powerful open-source image generation model — one model for both text-to-image generation and image editing. Its 7B-parameter visual component (32 single-stream DiT layers) sits behind a Qwen3-VL 8B text encoder and a 64-channel RGBA autoencoder that outputs native transparency. In MindCraft Studio it runs locally as an MLX model and stands out for bilingual (English / Chinese) prompt understanding, accurate text inside the image, and editing with up to 10 reference images.

This guide details how the model behaves inside MindCraft Studio and the prompting techniques that get the best results.

---

1. Recommended Settings

The model's native resolution is 2K, but on a Mac it runs at smaller sizes. MindCraft Studio's defaults are tuned for local hardware:

• Size — start at 512×512; step up to 1024×1024 for noticeably better results. Both sides must be a multiple of 16.
• Steps — the app defaults to 20. Use 40 (the model's own recommendation, roughly double the time) whenever fine detail or legible text matters. Low step counts are only for checking composition — never judge quality from them.
• Aspect ratio — pick one of the model's native ratios and render at a smaller size (see Aspect Ratios & Resolution). No reference image is needed for wide or panoramic shots — just set a wide size.
• No negative prompt — the negative-prompt field is hidden for this model. Never write "avoid X" or list what you don't want; describe only what should be in the frame.
• No strength slider — reference images are context, not a partially re-noised starting image. The instruction text is the only lever.

---

2. Writing Text-to-Image Prompts

• Write natural prose, not keywords. The model reads the prompt as a sentence, so grammar and explicit relationships ("in front of", "behind", "next to") materially change the result.
• Enrich short inputs with materials and textures, lighting, camera framing, and depth of field.
• Name the concrete details — a "85 mm portrait lens", "soft window light", "subsurface scattering on skin", "wet specular highlights on leaves". The model responds to concrete scene content, not tags like "photorealistic, 8k, masterpiece".
• Preserve diagram logic — for flowcharts or diagrams, keep the connections and arrows exactly as described.
• Phrase positively. Because there is no negative field, say what is in the frame, never what is not.

Worked example — short to expanded: Short: a corgi playing a ukulele at a picnic Expanded: A realistic photograph of a corgi sitting upright on a red-and-white checked picnic blanket in a sunny park, playing a wooden ukulele. The corgi looks up with a playful expression, its orange and white fur soft in bright midday sunlight. On the blanket around it sit a wicker basket, a thermos, and a few scattered daisies. Shot on an 85 mm lens with a shallow depth of field; the trees behind blur into green bokeh. Warm, natural color balance, gentle contrast.

---

3. Rendering Crisp Text & Typography

Typography is where this release improved most — lean into text-heavy work (posters, signs, packaging, infographics) rather than avoiding it:

• Put the exact wording in double quotes — a shop sign reading "MindCraft".
• Never translate, reword, or correct text inside the quotes — reproduce it exactly, capitalization and punctuation included.
• Style it in words — "MindCraft" in bold orange sans-serif at the top centre of the poster.
• Chinese text stays Chinese — and can name a typeface family (such as kai or song style) when the look matters.
• Quote the string, paraphrase the rest — only text that must render verbatim goes inside the quotes.

Example Prompt: A minimal poster for a jazz festival, a bold white title reading "JAZZ NIGHTS" across the top centre, deep navy background, abstract warm-gold instrument silhouettes along the lower half, soft spotlight from above.

---

4. Transparent Backgrounds (RGBA) — Use the α Toggle

Qwen-Image-2.1's VAE outputs a real alpha channel, so it can produce transparent PNGs natively. In MindCraft Studio it is the only model with a transparent-background toggle — the α button, tooltip "Transparent Background" — next to the prompt box.

• When you want a transparent background, a cut-out, a sticker, a PNG with no background, or an asset for compositing: switch on the α toggle. The same toggle also covers editing inside a transparent layer (changing a subject's expression while keeping the background transparent) and cutting a subject out of an ordinary photo — give a normal photo and ask for the subject alone on a transparent background. Just describe the subject in words and generate; the app handles everything else.
• The edge is soft at low step counts — raise the steps when you need a clean cut-out.
• A white border usually means "sticker". The word sticker makes the model draw a die-cut white margin, like a printed sticker. If you want no border, drop "sticker" and ask for the subject alone on a transparent background.

---

5. Prompting Image Edits

The reference image is read by the same encoder as the text, so write the instruction as a sentence about the picture, not as a command to a tool:

• Describe the outcome, not the operation. "A red fox sitting on a sandy beach at sunset" works; "change the background" asks for a procedure. Name the subject, the new context, and the lighting.
• Restate what must survive. Identity and pose hold up on their own, but naming them helps: "…, keeping the fox's face, pose, and fur exactly as they are."
• One change per instruction. Several unrelated whole-image changes at once (recolour the fur and swap the background and add text) is where results go muddy — do them in successive edits. Region-anchored changes are the exception (see Local editing).
• No negation — the same positive-only rule as text-to-image.
• Set the shape first. Content is easy to change, framing is not — turning a square photo into a wide panorama is the hard case, so set the wide output before editing.

Performance: an edit is far heavier than a text-to-image at the same size. Start edits at 256×256 with 20 steps. A 512×512 edit on a 24 GB machine can stall for a very long time instead of finishing — if the progress indicator stops moving, lower the size rather than waiting. Don't raise size and steps together.

---

6. Multi-Reference & Local Edits

• Multi-reference (up to 10 images). The model combines references: "put the person from the first image on the sofa from the second." Order matters — refer to images by their attachment order ("the first image", "the second image"). Each reference slows generation, so use the fewest that carry the information.
• Local editing — three ways to point at a region. The model natively understands regions drawn straight onto the image — Qwen upstream demonstrates this with circles and painted annotations, telling the model things like "the blue circle" or "the red area". MindCraft Studio surfaces the same ability as the Annotate tool in the Modify menu: draw ellipses, boxes, freehand strokes, arrows, or text in different colours right on the photo, then pick Generate (direct) mode so the annotated image goes straight to the model — no region description needed in words, though you may still refer to the annotation's colour ("the area I painted green") to steer it.
• Local editing — Inpaint & Eraser. You can also avoid describing the region entirely and just paint it: the Inpaint tool sends a marked-up image plus an instruction that names the colour, Only replace the red area(s) with <your content> to complete the image, keep other pixels unchanged. The Eraser tool paints a filled region and asks the model to fill it in.
• Regions named in text work too. Upstream's own example: "Remove the metal watch in the blue circle, change the hair in the red circle to black, and replace the area in the green circle with gray short-sleeved linen pajamas." Several region-anchored changes in one instruction are fine — naming each region keeps them from bleeding into each other.
• Mark only what must change. A very large marked area leaves the model less context to work with — if your change covers most of the frame, re-describing the whole image works better.

Showcase — panorama, infographic & storyboard (upstream's blog demonstrates these three with a single reference image; they're ordinary edits, so run them in the app):

• Panorama — set the wide output before editing (pick a wide ratio in the size menu, e.g. 16:9 at a small size), then attach the photo and prompt:

  Extend this photo into a wide panoramic view of the same mountain valley, continuing the ridge, sky, and morning light on both sides. Keep the hiker, his pose, and the foreground rocks exactly as they are. Match the original perspective and light direction so the seams blend seamlessly.

• Infographic — attach a product photo and ask for a detailed, information-rich composition:

  Turn this product photo into a detailed infographic poster with the product as the hero. Add an exploded view of the components, a bold title reading "AERO BOTTLE", labelled spec callouts with thin white leader lines, scale markings, and a clean technical layout on a light gray background.

• Storyboard — attach a character reference sheet and ask for a multi-panel sequence:

  Turn this three-view character sheet into a six-panel storyboard. Keep the same character design, face, and costume colors in every panel; each panel shows her in a different scene and lighting with consistent framing. Arrange the panels in a clean grid with even gutters.

Also worth trying: cut-out — attach a photo, ask for the subject alone on a transparent background (α toggle on).

---

7. Aspect Ratios & Resolution

Qwen-Image-2.1's native aspect ratios are 2K-resolution. On a Mac, choose the ratio and render at a smaller size (sides in multiples of 16):

---

8. Constraints & Negative Guidance

This model has no negative-prompt field, so control the output entirely through what you describe:

• Never phrase a request as a prohibition — "a table without chopsticks" becomes a description of the table alone, not a list of what to exclude.
• For faithful edits, restate what must survive once, affirmatively: "keep the rest of the image exactly as it is."
• For typography hygiene, quote the exact string that should appear and style it — there is no room left for gibberish.

---

Ready to generate? Learn about prompt configurations in The Prompt Editor or build multi-reference scenes on the Infinite Canvas.